#407 32X on-chip 2k rasterization code test (slower)

XProger · Apr 23, 2022 · fedd2ed · fedd2ed
1 parent 93058da
commit fedd2ed
Show file tree

Hide file tree

Showing 9 changed files with 94 additions and 6 deletions.
diff --git a/src/platform/32x/32x.h b/src/platform/32x/32x.h
@@ -151,4 +151,8 @@ extern "C"
     void CacheClearLine(void* ptr);
 }
 
+#define CacheClear()\
+    CacheControl(0);\
+    CacheControl(SH2_CCTL_CP | SH2_CCTL_CE);
+
 #endif
diff --git a/src/platform/32x/asm/common.i b/src/platform/32x/asm/common.i
@@ -9,12 +9,6 @@
 // int16 e10, e11, e12  // rotation
 // int16 e20, e21, e22  // rotation
 // int16 e03, e13, e23  // translation
-//
-// word [ N Z Z ]  // rot
-//      [ Z N Z ]  // rot
-//      [ Z Z N ]  // rot
-//      [ Z Z Z ]  // trans
-// long [ NZ ZZ NZ ZZ NZ ZZ ]
 
 // row[0] rotation
 #define M00     0

diff --git a/src/platform/32x/asm/matrixSetIdentity.s b/src/platform/32x/asm/matrixSetIdentity.s
@@ -5,6 +5,12 @@ SEG_MATH
 #define ZZ      r1
 #define NZ      r2
 
+// word [ N Z Z ]  // rot
+//      [ Z N Z ]  // rot
+//      [ Z Z N ]  // rot
+//      [ Z Z Z ]  // trans
+// long [ NZ ZZ NZ ZZ NZ ZZ ]
+
 .align 4
 .global _matrixSetIdentity_asm
 _matrixSetIdentity_asm:

diff --git a/src/platform/32x/asm/rasterize.s b/src/platform/32x/asm/rasterize.s
@@ -44,6 +44,14 @@ var_fb:
         // write per but allow transparent write for byte & word
         .long 0x24020200
 var_table:
+/* 2k on-chip test
+        .long 0xC0000000 + 516 + 416 + 256 + 18 //_rasterizeS_asm
+        .long 0xC0000000 + 516 + 416 + 18 //_rasterizeF_asm
+        .long 0xC0000000 + 516 + 18 //_rasterizeFT_asm
+        .long 0xC0000000 + 516 + 18 //_rasterizeFT_asm
+        .long 0xC0000000 + 20 //_rasterizeGT_asm
+        .long 0xC0000000 + 20 //_rasterizeGT_asm
+*/
         .long _rasterizeS_asm
         .long _rasterizeF_asm
         .long _rasterizeFT_asm

diff --git a/src/platform/32x/asm/rasterizeF.s b/src/platform/32x/asm/rasterizeF.s
@@ -30,6 +30,9 @@ SEG_RASTER
 #define LMAP    inv
 
 .align 4
+.global _rasterizeF_asm_start
+_rasterizeF_asm_start:
+
 .exit:
         // pop
         mov.l   @sp+, r14
@@ -213,3 +216,7 @@ var_LMAP_ADDR:
         .long   _gLightmap_base
 var_divTable:
         .long   _divTable
+
+.align 2
+.global _rasterizeF_asm_end
+_rasterizeF_asm_end:
diff --git a/src/platform/32x/asm/rasterizeFT.s b/src/platform/32x/asm/rasterizeFT.s
@@ -57,6 +57,9 @@ SP_R   = 24
 SP_SIZE = 28
 
 .align 4
+.global _rasterizeFT_asm_start
+_rasterizeFT_asm_start:
+
 .exit:
         // pop
         add     #SP_SIZE, sp
@@ -317,3 +320,7 @@ var_divTable:
         .long   _divTable
 var_gTile:
         .long   _gTile
+
+.align 2
+.global _rasterizeFT_asm_end
+_rasterizeFT_asm_end:
diff --git a/src/platform/32x/asm/rasterizeGT.s b/src/platform/32x/asm/rasterizeGT.s
@@ -69,6 +69,10 @@ SP_R   = 28
 SP_SIZE = 32
 
 .align 4
+
+.global _rasterizeGT_asm_start
+_rasterizeGT_asm_start:
+
 .exit:
         // pop
         add     #SP_SIZE, sp
@@ -400,3 +404,7 @@ var_divTable:
         .long   _divTable
 var_gTile:
         .long   _gTile
+
+.align 2
+.global _rasterizeGT_asm_end
+_rasterizeGT_asm_end:
diff --git a/src/platform/32x/asm/rasterizeS.s b/src/platform/32x/asm/rasterizeS.s
@@ -29,6 +29,9 @@ SEG_RASTER
 #define ih      inv
 
 .align 4
+.global _rasterizeS_asm_start
+_rasterizeS_asm_start:
+
 .exit:
         // pop
         mov.l   @sp+, r14
@@ -188,3 +191,7 @@ var_LMAP_ADDR:
         .long   _gLightmap_base
 var_divTable:
         .long   _divTable
+
+.align 2
+.global _rasterizeS_asm_end
+_rasterizeS_asm_end:
diff --git a/src/platform/32x/render.cpp b/src/platform/32x/render.cpp
@@ -685,6 +685,37 @@ void flush_c()
         return;
 
     gFacesBase = gFaces;
+/*
+    CacheControl(0);
+    CacheControl(SH2_CCTL_CP | SH2_CCTL_CE | SH2_CCTL_TW);
+
+    extern int32 rasterizeGT_asm_start;
+    extern int32 rasterizeGT_asm_end;
+
+    int32 size = intptr_t(&rasterizeGT_asm_end) - intptr_t(&rasterizeGT_asm_start);
+    fast_memcpy((void*)(0xC0000000 + 0), &rasterizeGT_asm_start, size >> 2); // 516
+
+    extern int32 rasterizeFT_asm_start;
+    extern int32 rasterizeFT_asm_end;
+
+    size = intptr_t(&rasterizeFT_asm_end) - intptr_t(&rasterizeFT_asm_start);
+    fast_memcpy((void*)(0xC0000000 + 516), &rasterizeFT_asm_start, size >> 2); // 416
+
+    extern int32 rasterizeF_asm_start;
+    extern int32 rasterizeF_asm_end;
+
+    size = intptr_t(&rasterizeF_asm_end) - intptr_t(&rasterizeF_asm_start);
+    fast_memcpy((void*)(0xC0000000 + 516 + 416), &rasterizeF_asm_start, size >> 2); // 256
+
+    extern int32 rasterizeS_asm_start;
+    extern int32 rasterizeS_asm_end;
+
+    size = intptr_t(&rasterizeS_asm_end) - intptr_t(&rasterizeS_asm_start);
+    fast_memcpy((void*)(0xC0000000 + 516 + 416 + 256), &rasterizeS_asm_start, size >> 2); // 224
+
+    //extern int32 fps;
+    //fps = size;
+*/
 
     PROFILE(CNT_FLUSH);
 
@@ -753,6 +784,10 @@ void flush_c()
 
         } while (face);
     }
+/*
+    CacheControl(0);
+    CacheControl(SH2_CCTL_CP | SH2_CCTL_CE);
+*/
 }
 #endif
 
@@ -843,6 +878,18 @@ void renderInit()
     gFacesBase = gFaces;
 }
 
+void renderFree()
+{
+}
+
+void renderLevelInit()
+{
+}
+
+void renderLevelFree()
+{
+}
+
 extern "C" X_NOINLINE void drawTriangle(uint32 flags, VertexLink* v)
 {
     VertexLink* v0 = v + 0;